SpIRL: Spatially-aware image representation learning under the supervision of relative position descriptors

Download PDF
Open Webpage

Logan Servant, Michaël Clément, Laurent Wendling, Camille Kurtz

Published: 2026, Last Modified: 10 Nov 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Semantic information holds back neural networks in their ability to create spatial representations.•New unsupervised strategies to learn spatially-aware image representation.•Relative position descriptors can be used to monitor the representation optimization from neural networks.•Proposition of a new synthetic image dataset not biased by semantic information.•Interest of fine-grained spatial representation for spatial relationship recognition and image retrieval.